70 likes | 166 Views
End-host Perspectives on Congestion Management. Murari Sridharan muraris@microsoft.com CONEX BOF, IETF 76, Hiroshima. Its all about the apps . Apps have diverse needs P2P, VoIP, TV, online apps, games Filling the pipe while being TCP fair
E N D
End-host Perspectives on Congestion Management Murari Sridharan muraris@microsoft.com CONEX BOF, IETF 76, Hiroshima
Its all about the apps • Apps have diverse needs • P2P, VoIP, TV, online apps, games • Filling the pipe while being TCP fair • High-speed congestion control, Autotuning the receive window, host OS (including app) bottlenecks • Low priority scavenger service • Protocols changes to reduce latency • Multipath transport to maximize link capacity, improve resiliency
End-host view of the network • Mostly works as expected • Establishing connectivity is sometimes problematic • Largely remains a black box • Congestion, bottleneck capacity is implicitly inferred • Network seems change/risk averse • Imposes complex usage requirements • Volume caps, pay-per-use, metered data plans …
Network’s view of the end-hosts • End-hosts cannot be trusted, need to be explicitly controlled • Application performance needs can be inferred and improved upon by inspecting packets on the wire • End-hosts typically establish connectivity to well-known servers • And the reality is …
Splitting at the seams • Devices, Devices, Devices … • IGDs/NATs/Firewalls/DPI devices/Load balancers/Soft switches • Layering violations, RFC non-conformance, no future proofing • Backwards compatibility • Thanks to deep packet inspection you can’t even touch padding bits! • Hampered ECN deployment • New protocols hard or impossible to deploy • No changes to protocol on the wire • Takes time for apps to move to using new APIs • Not all apps need TCP, but only TCP passes reliably E2E • Tunneling, feedback loops & TCP over TCP issues • Network traffic management is a nightmare due to increased appetite for large volumes
CONEX: Things to consider • Let’s please stop inferring each other. • How can the end-host learn about network policy? • Explicit feedback to/from the network • OS has tons of context than the small amount reflected in the packets • Wireless/Cellular network operators are moving towards usage based models • Tactically, policy based routing/interface selection is the high-order bit • How does the end-host learn about the network policy?
CONEX: Things to consider • Incremental deployment & incentives • Bottleneck is largely within the home and/or last mile • Congestion exposure doesn’t add any benefit but shouldn’t make it worse • Primal-dual congestion control • Performance is a function of source rate controller, AQM, marking thresholds, congestion measure • Measure of congestion, feedback of congestion measure, and use of congestion measure by the source controller are all important. • Beyond TCP – UDP(DCCP) + ECN